Backtest Issues





Kerry Back

Issues

  • Should loop
    • In practice, we would use all past data to train a model before trading on it
    • In a backtest, we should follow the same procedure: at each date us all past data to train and then predict
  • Should include industry as a feature
    • E.g., buy value industries or best value in each industry?
    • Parameters may vary by industry
  • Feature and target transformations

Transforming cross-sections

  • (time, stock) data is called panel data
  • The set of stocks at a point in time is called a cross-section.
  • Should transform and standardize each cross-section
    • Relative rankings of features at a point in time should matter rather than ranking across time
    • To “beat the market,” we need to predict which stocks will beat others not what the overall market will do